With advancements in IoT and edge computing, multi-tenant environments have emerged as a paradigm where multiple users efficiently execute tasks using shared resources. However, ensuring data privacy among tenants and addressing the complexity of dynamic network traffic pose challenges for conventional anomaly detection systems. In this study, we propose a federated learning-based network anomaly detection system that integrates autoencoder models based on Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The LSTM autoencoder consistently exhibited stable performance, while the GRU autoencoder demonstrated superior computational efficiency, making the system well-suited for resource-constrained environments. To preserve data privacy, each tenant independently trains on local data, sharing only model parameters with a central server to construct a global model. Experimental evaluations using state-of-the-art malicious traffic datasets-Botnet, Redline Stealer, Snake Key-Logger, Matanbuchus, and Astaroth Guildma-demonstrate high accuracy, precision, and low false positive rates. Notably, our system outperforms traditional methods such as Principal Component Analysis (PCA) and Support Vector Machines (SVM), achieving significant performance improvements. These results confirm that our proposed system provides a robust solution for detecting anomalous traffic in multi-tenant environments, ensuring both high detection performance and strict data privacy.